SFT Conflicts, RL Coexists: A Theoretical and Empirical Analysis of Multi-Task Learning for LLMs

Benchmark Model Rank Results
mathematical-reasoning-on-math-500Adapted Parallel-RL (DeepSeek-R1-Distill-Qwen-7B)#19Accuracy: 96.2
mathematical-reasoning-on-aime-2025Adapted Parallel-RL (DeepSeek-R1-Distill-Qwen-7B)#76Accuracy: 59.0
multiple-choice-question-answering-mcqa-on-gpqa-diamondAdapted Parallel-RL (DeepSeek-R1-Distill-Qwen-7B)#108Accuracy: 54.0